Mutually Assured Compute

Mutually assured compute is a proposed deterrence regime in which multiple nations holding frontier AI compute agree that any unilateral attempt to develop superintelligence will trigger retaliatory destruction of compute infrastructure across all signatories. The mechanism generalizes the logic of nuclear mutually assured destruction to the domain of AI hardware, with each side's ability to destroy the other's training clusters serving as the deterrent against a first strike on the agreement.

The proposal rests on three technical prerequisites. First, attribution: each signatory must be able to identify when another is conducting a forbidden training run, which requires either chip-level monitoring, energy-grid signatures, or supply-chain tracking. Second, verification: a mechanism, ideally inference-only, that confirms no party is secretly scaling toward superintelligence while publicly complying. Third, response capacity: each signatory must retain the ability to destroy the other's compute fast enough that a first strike cannot complete a superintelligence project before retaliation lands. The proposal treats compute as the AI equivalent of nuclear arsenals, and the agreement as a non-proliferation treaty enforced by retaliatory capability rather than inspection.

The benefit is that the regime could in principle slow the race to superintelligence without requiring any party to disarm unilaterally, since the deterrent is symmetric and each side retains its full stack. The cost is that it depends on attribution and verification technologies that do not yet exist at the required fidelity, and on the willingness of nations to destroy billions of dollars of allied infrastructure in response to a treaty violation. The regime is also vulnerable to covert projects that escape detection, and to actors who are willing to absorb a retaliatory strike because they believe they can complete superintelligence fast enough to win anyway.

Open questions include whether attribution can ever be reliable enough to detect covert frontier runs, whether the regime would deter or merely accelerate a hidden race, and whether a sufficiently capable AI system could subvert the verification infrastructure itself. Critics also note that mutually assured destruction worked for nuclear weapons largely because the consequences of use were unambiguous, whereas the threshold for 'superintelligence' is contested and could be reached through incremental capability gains rather than a single detectable training run.